Electronic Music Artist Identification

TitleElectronic Music Artist Identification
Publication TypeMaster Thesis
Year of Publication2012
AuthorsMelidis, P.
AbstractWe address and research on the problem of Electronic Music Artist Identification. On an in-house built dataset, we computed a variety of different descriptors (MFCC, HPCP etc) and applied machine learning techniques (SMO, ADTrees) in order to discover those distinct (stylistic and aesthetic) characteristics that are associated with Electronic Music producers. We used 111 Artists in the dataset, amounting to a total of 5949 songs. Using 5-Fold cross validation, we yielded an accuracy of 21.3649 % when using Zipfian MFCC Coded Words, and 19.4109 % with normal MFCC descriptors. When using an independent hold-out dataset for validation, results were -as expected- worser (6.66% accuracy, when using Spectral Features and 4.74% when using MFCCs). We further investigate the effect of different feature sets on the overall performance, and also the effect of the artist number on the accuracy of our system. We furtherly conclude that there are certain specific elements among Electronic Music artists that make the identification task easier (e.g. occurrence of sample-based techniques)